OpenAI's image input requirements are specific enough to design a pipeline around. Supported files are PNG, JPEG, WEBP and non-animated GIF. A request can carry up to 512 MB of total payload and up to 1,500 images. Per image, the API applies up to 30,000 patches after the model's own resize rules, and that limit is the one people trip over: images that exceed it are rejected outright, not quietly shrunk to fit. Then detail levels decide how much of the image survives. Low fits everything inside 512x512. High fits within 2,048x2,048 and 2,500 patches. Original preserves your dimensions but scales down anything larger than 65,535 px on a side, and on gpt-5.5 the original ceiling drops to 10,000 patches and 6,000 px. Practically: if you send a 4,000 px screenshot at high detail, the model throws most of it away. Sending it at 2,048 px on the long edge gives the model the same tokens, in a smaller file, with a much shorter upload. Skwish resizes, converts and compresses in the browser, keeps a per-file before and after readout, and packs a batch into one ZIP. No image leaves your device.